Embed Size (px)
Citation preview
Improving Window Switching Interfaces
Susanne Tak1, Andy Cockburn1, Keith Humm1, David Ahlström2, Carl Gutwin3, and Joey Scarr1
1 Computer Science and Software Engineering, University of Canterbury,
Private Bag 4800, Christchurch 8140, New Zealand
[email protected], [email protected],
{khu23,jls129}@student.canterbury.ac.nz 2 Department of Informatics Systems, Klagenfurt University,
Universitätsstrasse 65-67, A-9020 Klagenfurt, Austria
[email protected] 3 Department of Computer Science, University of Saskatchewan,
110 Science Place, Saskatoon, Saskatchewan, S7N 5C9, Canada
Abstract. Switching between windows on a computer is a frequent activity, but
current switching mechanisms make it difficult to find items. We carried out a
longitudinal study that recorded actual window switching behaviour. We found
that window revisitation is very common, and that people spend most time
working with a small set of windows and applications. We identify two design
principles from these observations. First, spatial constancy in the layout of
items in a switching interface can aid memorability and support revisitation.
Second, gradually adjusting the size of application and window zones in a
switcher can improve visibility and targeting for frequently-used items. We
carried out two studies to confirm the value of these design ideas. The first
showed that spatially stable layouts are significantly faster than the commonly-
used recency layout. The second showed that gradual adjustments to
accommodate new applications and windows do not reduce performance.
Keywords: window switching, revisitation patterns, spatial constancy
1 Introduction
Switching between windows like email applications, word processors, and Web
browsers is a very common task. A previous study [1] found that the mean time
between window switches is only 20.9 seconds and that users have more than eight
windows open more than 78% of the time. It has also been reported that the average
number of simultaneously opened windows increases with available display space:
from four for single monitor users to up to 18 for users with multiple monitors [2].
Current interface methods for window switching have changed relatively little
since early graphical user interfaces – clicking on a window brings it into focus, as
does selecting the window from a spatial iconic representation (e.g., the Windows
Taskbar) or from a recency list (e.g., the Windows Alt+Tab display). Recent research
and commercial systems demonstrate alternatives to these mechanisms, but problems
still exist – with more than a few windows in the display, users must often carry out a
laborious search to find the desired window, even if that window is used frequently.
Despite the different designs that have been proposed, no window switching tools
are based on empirical evidence about how people revisit windows in actual desktop
work. To address this limitation, and to identify new design principles for window
switching tools, we carried out three studies. We first conducted a longitudinal study
where window switching behaviour was recorded; this study showed that revisitation
is frequent, both to applications and to specific windows, and that most switches are
to a small number of applications. From this study, we identified the design principles
of spatial constancy and morphing target sizes. Spatial constancy supports window
revisitation by keeping thumbnails for activating windows and applications in the
same place in the switching display, allowing users to build up spatial memory of
frequently-used items. Morphing target sizes allocates more space to frequently-used
applications and windows, improving Fitts’ Law targeting time for the most frequent
items, and allowing for the addition of new items.
Our second and third studies tested the value of these two principles. The second
showed that spatial constancy is effective: stable layouts are significantly faster than
recency layouts (similar to Windows Alt+Tab). The third study showed that gradual
adjustments in the display’s layout do not reduce (or improve) performance. We
discuss how these successful results can be deployed in new window switching tools.
2 Related Work
Two areas of related work inform our investigation: research on task and window
switching interfaces; and studies of user behaviour with desktops and windows.
2.1 Window Switching Interfaces
The importance and frequency of window switching has led to extensive research and
development into improved interfaces for the task. Two interface properties help
distinguish window switching approaches: first, the type of information used to form
and display relationships between windows, such as temporal, spatial, or semantic
data about the windows; and second, the degree to which systems try to automatically
establish these relationships, with some being entirely manual while others use
sophisticated predictive methods to automatically establish window relationships.
Henderson and Card’s seminal work with the Rooms [3] virtual desktop manager
was almost entirely manual, with the user assuming all responsibility for the spatial
placement of windows within a room based metaphor. Scalable Fabric [4] also uses
manual relationship controls, including extensive support for zooming, but unlike
Rooms it provides little explicit structure for grouping windows. The default desktop
behaviour of Microsoft Windows XP/Vista is also largely manual (users place
windows where they wish), although it automatically groups windows belonging to
each application in the Taskbar. Microsoft’s GroupBar [2] maintains manual control,
but replaces the Taskbar’s application-based grouping with user-defined task groups.
The primary limitation with manual control of window relationships is that users
must carry out explicit actions to gain potential benefits. To remove the dependence
on manual actions many systems automatically form window relationships. The most
basic and widely deployed automatic relationship system is the familiar Alt+Tab key
binding in Microsoft Windows operating systems. Alt+Tab allows users to rapidly
flip through a temporally based ‘z-ordering’ of windows on the display. Kumar et al.
[5] observe that Alt+Tab is very efficient when the number of windows is low, but
researchers have also labelled the method ‘tedious’ [6]. Mac OS X’s Exposé also uses
an automatic layout, displaying thumbnails of all windows at once; but the layout is
not spatially constant, so users cannot accurately predict where specific windows will
be located, demanding visual search to find them.
SWISH [7], UMEA [8], TaskTracer [9], RelAltTab [10] and Taskposé [11] all use
sophisticated methods to automatically adapt to user activities. SWISH uses temporal
relationships between window focus events as well as window titles to establish
semantic relationships. Their evaluations suggested 70% accuracy rates in assigning
windows to task groups. Similarly, TaskTracer uses machine learning to modify the
Microsoft Start menu, Taskbar and Windows Explorer. RelAltTab uses similar
methods to modify the Alt+Tab window order. Taskposé provides an overview in
which windows drift towards each other dependent on their temporal relationships,
based on the results of a ‘WindowRank’ algorithm. Taskposé windows also gradually
enlarge to reflect their relative importance as calculated by the algorithm.
The primary limitations of automatically adaptive systems are that they can
incorrectly predict the user’s intention and that users can fail to understand or
anticipate the system’s adaptation [12]. When this happens users must resort to time-
consuming visual search of candidate targets. Of all the previous designs, however,
only two – WindowScape [13] and Elastic Windows [14] – used stable layouts to help
improve memorability of previously-used windows.
2.2 Studies of Window Use
Gaylin [15] provides an analysis of window activities from the early days of graphical
user interfaces, based on 22 minute observations of nine participants. Observations
show that window switching activities were far more frequent than window creation,
deletion, or geometry management. Hutchings et al. [1] update these findings using
automatic logs of 39 participants over a 3-week period. They report on how window
management activities differ across single- and multi-monitor display use (in common
with Grudin’s earlier field study [6]) but their data also highlights important general
characteristics of window management. This includes the finding that window
switching is extremely frequent, with a mean window activation time of 20.9 seconds,
and a median of only 3.77 seconds. This frenetic frequency of window switching is
confirmed by Mackinlay and Royer [16], who also conducted a log analysis of
window switching. Hutching et al.’s data also shows that users normally have many
windows open, with eight or more windows open 78.1% of the time.
Many types of human behaviour are highly repetitive, as observed by Zipf’s Law
[17] and the Pareto Principle [18] (also called the “80-20” rule), where a large portion
of effects comes from a small portion of causes. Zipfian distributions have been
observed in many areas of computer use [19], such as frequency of command use and
menu use. Although window and application switching activities are among the most
frequent in computer use, with one study showing a mean time of 21 seconds between
actions [1], we are unaware of any empirical studies of users’ patterns of revisitation
to windows and applications. To address this limitation, we carried out the
longitudinal study described in the following section.
3 Log Study of Application and Window Switching
To find out how users switch between applications and windows, we carried out a
longitudinal study that recorded windowing behaviour as users went about their
everyday tasks. We developed logging software for Windows XP that unobtrusively
monitors window switches, window creation/destruction events, changes in window
geometry, and the method used to switch windows (e.g., Alt+Tab, Taskbar, and direct
mouse click). Nine frequent computer users (18 to 55 years old) took part in a study
during which we recorded between 8 and 117 days of data per participant. Four
participants used dual monitors; five, single monitors (see Table 1).
Overall, we obtained 241 person-days of data. Only manual window switches were
included in the analysis: automatic switches (such as window/dialog pop-ups) were
removed from the data. This left a total of 45,377 switch events.
3.1 Results
Results are divided into three categories. First, we present data describing revisitation
patterns for the windows used each day. This characterises the users’ main activities
in window switching. Second, we describe long term revisitation patterns to
applications. While many windows are transient, existing only for immediate work
requirements (such as a window containing an email message during its composition)
applications are relatively stable. For many tasks, such as “search on the Web for
topic X” or “check my email inbox” the user’s target is likely to be the application
(e.g., Firefox or Outlook) rather than a specific window, so an application based
analysis is potentially informative. Third, we analyse the interface mechanisms used
to carry out window switching activities to gain insights to how current interfaces are
used and to determine whether users adopt similar or divergent patterns of behaviour.
Daily Window Revisitation. For each participant on each day, we analysed how
frequently each window was revisited. This was conducted by forming a ranked order
of windows according to their percentage of total daily switches.
The number of window switches per day for participants ranged from 5 to 807,
with a cross participant mean of 219 per day (s.d. 91). The number of distinct
windows switched to per day ranged from 3 to 177, with a cross participant mean of
39 (s.d. 19). In their study of Web revisitation, Tauscher and Greenberg [20] define
the recurrence rate R as the probability that any URL visit is a repeat of a previous
visit, giving R=(total URLs visited – distinct URLs visited)/total URLs visited ×100.
Adapting this formula for window revisitation gives a mean recurrence rate of 82%;
much higher than the 58-61% rate reported for the Web. This data shows that window
revisitation is a very common activity.
Analysing the same data with respect to the Pareto principle shows that 80% of
window switches were triggered by between 24 and 40% of windows for the different
participants (see Table 1), with a mean of 35.1%.
Table 1. Application and window switching data for the nine participants.
Pareto. 80% of
switches by what %
Interface %
P. single/
dual
# days #
switch
# apps Windows Apps Click Task-
bar
Alt+
Tab
1 single 10 1396 40 40 15 23.4 76.7 0.0
2 dual 117 17088 45 34 7 78.4 21.4 0.2
3 single 23 2588 23 37 17 37.0 47.7 15.2
4 single 16 3019 49 37 16 35.3 61.3 3.4
5 dual 9 2454 37 40 22 64.5 35.5 0.0
6 dual 30 10373 54 31 13 64.1 35.5 0.4
7 dual 8 2922 34 35 18 82.7 14.5 2.8
8 single 13 2989 39 24 18 22.7 75.3 1.9
9 single 15 2548 23 37 17 7.6 7.7 84.7
Revisitation to Applications across Sessions. As mentioned above, many user tasks
involve targeting an application rather than a specific window. For example, a user
may have several Web browser windows concurrently open, any of which can be used
for a pressing search task; or the user may need to target a calculator, file explorer,
etc. In each case the application is the target, not a specific window. We therefore
analysed each participant’s revisitation to applications for the duration of the study.
The number of distinct applications used by the participants ranged from 23 to 54,
with a mean of 38 (s.d. 10.6). For each participant the total number of switches to
each application was calculated, ranked by frequency, and converted to a percentage
of their total application switches. Fig. 1 shows this data for each participant. It is
clear that a few applications are used a lot, and a lot are used relatively little.
0%
5%
10%
15%
20%
25%
30%
35%
40%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
n th most targeted application
Pe
rce
nta
ge
of
sw
itc
he
s
1 2 3 4 5 6 7 8 9
Fig. 1. Percentage of switches to the 20 most frequent applications for each participant.
Table 1 shows Pareto principle data, revealing that most users’ application
revisitation roughly adheres to the 80-20 rule. For example, 22% of participant five’s
applications accounted for 80% of switches. All other participants’ application
revisitation was more pronounced than the Pareto principle predicts: e.g., only 7% of
participant two’s application switches accounted for 80% of switches.
3.3 Interface Mechanisms Used to Switch Windows
We analysed the main three interface mechanisms currently used to visit and revisit
windows: direct window clicks, selection from the Taskbar, and Alt+Tab. This
analysis was conducted to determine whether users made similar or divergent use of
the tools available. Table 1 shows each participant’s use of these tools. Two important
observations from this data are as follows. First, there are substantial differences
between participants in their use of Alt+Tab. Seven of the participants used it very
lightly (less than 3.5% of window activations) or not at all; one used it fairly often
(15.2%); and one used it almost exclusively (84.7%). The two participants who used
Alt+Tab heavily had a single monitor, suggesting that it might be most valuable for
users with constrained screen real-estate (see Fig. 2). Second, direct clicks and
Taskbar use were also influenced by screen real-estate, with dual monitor participants
using direct clicks more and the Taskbar less than single monitor participants.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
Direct Click Taskbar Alt-Tab
Mean
perc
en
tag
e o
f to
tal
sw
itch
es
Single monitor
Dual monitor
Fig. 2. Cross participant means of the percentage of window switches activated by clicking on
the window, Taskbar selections, and Alt+Tab. Error bars show ±1 standard error.
4 Design Principles for Window Switchers
The patterns observed in the log study strongly suggest that supporting revisitation
should be a main design goal in window switching tools. More specifically, this
support should allow users to quickly find and distinguish between previously-visited
windows in a switcher’s display. There are several possible ways to provide this
support (for example, previously-visited windows could be highlighted in the display,
or the most frequently-used windows could be shown first), but previous research in
psychology argues for an approach that makes use of spatial constancy, discussed in
Section 4.1. However, user’s patterns of behaviour change over time: for instance, a
user may replace an application with a different vendor’s system. Complete spatial
constancy does not allow for the addition/removal of items. We address this problem
with the use of size morphing (Section 4.2). Morphing target sizes allow for the
addition of new items while maintaining as much spatial stability as possible. Also,
allocating more space to frequently-used applications and windows will reduce their
Fitts’ Law targeting times. Given that some applications and windows are used much
more frequently than others this might increase overall performance.
4.1 Supporting Window and Application Revisitation with Spatial Constancy
Spatial location memory is a person’s memory of where objects are in space. It is well
developed and can be extremely fast: studies have shown that items can be found in
time proportional to the logarithm of the size of the set, which can be much faster than
the linear search time needed for unorganised sets [19]. This fast performance is
enabled by spatial constancy – that is, items remaining in the same location over time.
This idea has been known since the first interface design guidelines [21], and many
studies have demonstrated its effectiveness.
There is also evidence that spatial location memory in user interfaces is
surprisingly robust. In studies of the Data Mountain’s spatial layout of thumbnail
images [22] users were able to quickly and accurately recall the location of specific
targets four months after originally creating a layout of 100 images. Furthermore,
their retrieval performance was not significantly harmed when the images were
replaced with blank outlines. Spatial constancy, however, has received little attention
in research on task- and window switching tools (with a few exceptions [13, 14]).
Applying the idea of spatial constancy in a window switcher implies that windows
and applications should not move in the switcher’s display. Within a work session,
this design approach has clear advantages: since revisitation is strong, people will
quickly learn the locations of the most frequently-used windows. Spatial stability
offers similar advantages across work sessions. Since people regularly return to a
relatively small number of applications, the problem of where to place individual
windows in the switcher display can be solved by creating ‘application zones’ that are
themselves spatially constant, and that reflect people’s longer-term repeating work
patterns. Spatial constancy can thus be applied in a hierarchical fashion: applications
are given stable zones in the switcher display, since these change slowly over the long
term; and within each zone, the application windows used in the current work session
are placed in stable spatial locations. The application-based organisation provides an
initial guide to the location of a new window, but as the window is used more and
more frequently, users will start to remember its location as a separate entity.
4.2 Size Morphing to Accommodate Change and Optimise Performance
The design principle of spatial constancy potentially conflicts with changes in patterns
of behaviour. For example, when users replace one application with another how can
spatial stability be maintained? Also, given that some applications and windows are
used much more frequently than others, how can the switching interface depict the
relative importance of applications and windows, and optimise the acquisition of
frequent targets?
Our proposed solution is to use gradual size ‘morphing’ to adjust the sizes of
application zones and window thumbnails. Morphing avoids abrupt changes in layout
that would damage spatial memory, while allowing the introduction of new items and
enabling frequent targets to be enlarged to enhance their visibility and to reduce
pointing time.
Fig. 3 shows a mock-up of our design, which uses the entire screen when activated.
Clicking on a zone or thumbnail immediately switches to the associated application or
window. Each zone contains thumbnails representing all windows associated with the
application, scaled and tiled to fit. Initially all application zones are of equal size (see
Fig. 3a), but they gradually morph in size to reflect frequency of use (see Fig. 3b).
(a) Initial view. (b) After morphing.
Fig. 3. Prototype application zones, before and after size morphing. The application and
window in the top left have been used most frequently.
5 Evaluating Spatial Constancy and Morphing
Our design principles assume that users will perform better at switching to windows
when thumbnails remain in spatially stable locations than when they are rearranged
according to recency or other properties. It also assumes that size morphing will allow
users to maintain their spatial memory while the interface adapts. We conducted two
experiments to separately investigate these issues.
5.1 Experiment 1: Spatial, Recency, or Frequency Ordering?
A window switching interface can order items in a variety of ways, including spatially
or by recency (like Alt+Tab). Other orders are also possible: frequency reordering
might work well for Zipf-like distributions; random orders may perform well if visual
popout effects are strong. This experiment, therefore, investigates the performance
impact of four different orderings (spatially stable, recency order, frequency order,
and random order) for tasks involving acquisition of targets in a Zipf-like distribution.
The experimental interface consisted of a grid of distinct icons (Fig. 4a) with a
cued target on the right. Participants were instructed to click on the target icon region
as quickly and accurately as possible, with each successful acquisition immediately
cueing the next. The spatially stable layout used an arbitrarily but stable order (i.e.,
icons never moved). In the random layout all items were randomly repositioned after
each selection. The recency reordering condition moved the most recently selected
item to the top left of the grid, pushing earlier items along in row-major order (similar
to Alt+Tab). Finally, frequency reordering repositioned items according to their
cumulative selection counts (most frequent at top left, in row-major order). Note that
frequency reordering rapidly stabilizes with a Zipf-like distribution of targets, while
recency ordering is less stable.
(a) Experiment 1. (b) Experiment 2 (post morphing).
Fig. 4. Experimental interfaces showing 16 icon regions and the target cue on the right.
To understand how performance with these interfaces is influenced by number of
targets, participants completed trials with 4, 9, 16, 25, 36, 49 and 64-item grids. The
Zipfian distribution of targets was generated by randomly selecting eight targets from
among the candidates (all four for the 4 item grid): one was cued 10 times, one 5
times, then 3, 2, 2, 1, 1, and 1 for the others.
Twenty-six students volunteered for the experiment (16 male, 10 female, 16-44
years old). Their participation lasted approximately 45 minutes.
Results and Discussion. The selection time dependent measure was analysed using a
4×3×7 RM-ANOVA for factors layout (stable, frequency, recency, random),
experience (novice, intermediate or expert) and items (4-64). Experience was
determined by assigning first-time icon selections as novice, 2nd-7th selections as
intermediate, and 8th-10th as expert.
All factors showed significant main effects: layout (F3,60=72, p<.001), experience
(F2,40=409, p<.001) and items (F6,120=135, p<.001) (see Fig. 5). Stable layouts were
the fastest (mean 1.2s) followed by frequency reordering (1.3s), recency reordering
(1.3s), and random (1.8s). Post hoc comparisons show pairwise differences between
all layouts except frequency and recency reordering, and frequency reordering and
stable (Bonferroni, p<.05). Fig. 5a shows a significant layout×experience interaction
(F6,120=47, p<.001) caused by relatively constant performance across experience with
random layouts in contrast to marked improvement with other layouts. Fig. 5b shows
a significant layout×items interaction (F18,360=14, p<.001), caused by the random
layout worsening much more rapidly across increased number of items than the other
three layouts. The results show that spatial constancy is beneficial. The stable,
frequency, and recency layouts all supported expertise development; random did not.
The recency layout (similar to Alt+Tab order) was outperformed by the stable layout.
0.00
0.50
1.00
1.50
2.00
2.50
3.00
Novice Intermetiate Expert
Mean time (secs)
RandomRecencyFrequencyStable
0.00
0.50
1.00
1.50
2.00
2.50
3.00
4 9 16 25 36 49 64 (a) Experience. (b) Number of items.
Fig. 5. Experiment 1 mean selection times (±1 standard error) with the four layouts.
5.2 Experiment 2: The Effect of Size Morphing
The second experiment examined the performance impact of the morphing behaviour
used for two purposes: to allow new windows/applications to be introduced, and to
change target sizes in response to the Zipf-like frequency distribution. Experiment 1
suggests that total stability is the ‘gold standard’, but total stability would prohibit the
addition of zones for new windows/applications, as well as prohibiting morphing size
adaptation to reduce the Fitts’ Law targeting time.
We therefore experimentally compared target acquisition performance using totally
stable placement (the gold standard control condition) as well as morphing behaviour
implemented with squarified [23] and spiral [24] treemaps. Treemaps recursively
divide 2D spaces into rectangles of various sizes, with size representing an underlying
quantitative data attribute. Various algorithms for creating treemaps exist, and two
main properties are the average aspect ratio and spatial stability (see [24] for a recent
review). Squarified and spiral treemaps were used in this experiment because they
respectively support low and high spatial stability. Fig. 4b shows the spiral
experimental condition after morphing. The experimental method, procedure and
design were similar to Experiment 1. There were seventeen participants (fifteen male,
two female, 21-35 years old).
Results and Discussion. Task time was analysed using a 3×3×7 RM-ANOVA for
factors layout, experience, and items. There were significant main effects for items
(F6,96=204, p<.001) and experience (F2,32=444, p<.001), but not for layout (F2,32=1,
p=.3). Mean times for the three layouts were similar at 1.3s, 1.4s and 1.4s with stable,
squarified and spiral respectively.
The absence of a significant difference between layouts is interesting and warrants
further analysis. The stable interface is not a practical solution to window switching
because it prohibits the addition of new applications and windows. However, total
stability best supports users in exploiting their spatial memory, so it is interesting that
it did not significantly outperform either of the two treemap layouts. It is reasonable
to suspect that the treemap interfaces allowed users to benefit from reduced Fitts’ Law
targeting performance with frequently selected items (which gradually morphed to
larger sizes), and that this counteracted the penalties associated with reduced absolute
stability, but this explanation is not supported by a significant layout×experience
interaction (F4,64=2, p=.1). More powerful experimental analysis is required.
6 Discussion
The studies provide important findings for the design of window switching tools:
• Within a work session, people frequently revisit previously-used windows;
• Revisitation to a small set of applications is also strong across sessions;
• Most users rely heavily on mouse-based window switching methods, but some
use keyboard methods almost exclusively;
• Spatially-constant item layouts are significantly faster than recency-based
layouts for a Zipfian distribution of target stimuli;
• Gradual size ‘morphing’ of item areas in the display did not affect
performance – importantly, morphing did not substantially harm the
participant’s spatial memory for target locations.
In the following paragraphs, we consider reasons why these design principles
succeeded and discuss design issues for real-world window switching tools.
6.1 Why the Principles Succeeded
The results of Experiment 1 add to previous findings showing performance
advantages for spatial constancy in comparison to alternatives. Furthermore, these are
the first results (that we know of) comparing spatially stable 2D layouts with
frequency and recency based layouts.
The stable layout was significantly faster than the recency layout, with
performance benefits increasing with expertise. This is easily explained – after each
selection in the recency layout users must either calculate the new position of their
targets or visually search for them, both of which demand time.
Performance with stable and frequency layouts was similar to each other, raising
the question of whether window switching interfaces should use frequency ordering.
However, the experiment used a Zipfian distribution of stimuli, causing the frequency
layout to quickly settle to spatial stability, so we believe that the frequency layout’s
comparative success is also explained by its spatial stability (after a short period of
reorganisation). In practical use, a frequency layout for windows is unlikely to
succeed due to the transient nature of most windows. Consequently, some other basis
for organisation is required, such as application zones. Although application zones
could be placed in a frequency layout, doing so would create placement instability
during early use, which might cause users to discard the system due to its
unpredictable behaviour. We therefore believe that using spatial stability as the main
layout principle is a preferable solution. Another reason for believing application
zones to be a useful placement strategy is that it quickly narrows the user’s search
space when there are several candidate windows from the same application – rather
than having to search the whole display, users can quickly dismiss the majority of
candidates by homing in on only those in the appropriate application zone.
The finding from the second experiment that morphing was no worse than absolute
spatial stability is important because it suggests that gradual layout changes allow
users to successfully use spatial memory while the display adapts to changes in their
behaviour. Although we have not yet tested item addition and deletion, we believe
that the spiral treemap algorithm used in Experiment 2 will allow users to maintain
their spatial understanding of the layout. We will test this in future work. We will also
model and test the Fitts’ Law performance advantages gained from morphing.
6.2 Design Considerations for Real-world Switching Tools
The log study and experiments suggest that our design principles can be successful,
but several design issues arise in translating these findings into a real-world switching
tool, as discussed below.
Number of windows and applications. Fig. 3 portrays our design intention, showing
four application zones and seven windows. However, the log study revealed that users
actually work with dozens of applications and windows. Can our design scale? We
are confident it can for several reasons. First, prior work (e.g., [22]) shows that users
can successfully learn and remember the locations of hundreds of targets. Second, our
experiments included up to 64 targets, and the benefits of spatial stability increased
with the number of targets. Third, our design uses the full display space to show and
access all windows and applications at once, allowing much larger targets than is
possible with visually compact tools such as the Windows Taskbar.
Flexibility in input device and interface behaviour. The log study showed that most
users relied on mouse input for window switching, but that one user used the Alt+Tab
key combination almost exclusively. Convenience will clearly influence the choice of
input device (e.g., using the mouse when a coffee cup is held in the other hand, or
using Alt+Tab when typing to avoid repositioning the hands). However, Alt+Tab is
also powerful when the user needs to ‘flip’ between a few recent windows. It is
therefore desirable that next generation window switching tools continue to support
flexibility in input devices and recency-based traversal. Our design can easily
accommodate both. For example, the tool could be activated by a key combination
(such as Alt+Tab) or by a dedicated mouse button or wheel; and the traditional
Alt+Tab recency list can be traversed by highlighting items on subsequent key
combination or button/wheel press.
Integrated support for application launch and window switching. Another potential
advantage of our design (yet to be evaluated) is that it integrates support for different
types of window switching tasks. In current interfaces, application launch facilities
are largely partitioned from window switching tools, yet the user activities are closely
related. For example, to “search on the Web for topic X” the user needs to acquire a
browser window. If the user starts by searching the Taskbar they will be unsuccessful
if no browser windows are active, necessitating a second action such as navigating
through the Start menu hierarchy to launch the browser. Alternatively, if the user
begins by launching the application, they will often gain a superfluous window (when
others were already available) adding to their window management load. Our design,
in contrast, provides a single interface mechanism for all window and application
activities: if the application zone is empty, the user clicks to launch; but if windows
are already available, any candidate can be immediately brought into focus.
Tailored layouts. Finally, the design could be adapted to allow users various
manual controls, such as ensuring that certain application zones are always displayed
in specific locations (perhaps to ensure consistency between a desktop and laptop
display), or excluding particular applications from appearing in the layout.
7 Conclusions and Future Work
Log analysis of real window and application switching showed that revisitation
(returning to previously used windows and applications) is an extremely frequent
activity in computer use. It also showed that most users activate their window and
application switches using the mouse as the input device, but that some users rely on
keyboard methods almost exclusively. These empirical characterisations of window
and application switching are new, and although they may confirm (or refute)
designers’ intuitions, empirical confirmation is necessary for informed design. The
main design message of the log analysis is that window and application switching
interfaces should improve efficiency by explicitly supporting revisitation.
We proposed two design principles to better support revisitation: spatial constancy
of the interface controls to activate applications and the windows belonging to them;
and size morphing to allow the spatial display to adapt to changes of behaviour and to
optimise selection of frequent targets. Two empirical studies showed, first, that
spatially stable layouts allow faster acquisition than recency and random layouts for
skewed distributions such as those occurring in window switching tasks, and second,
that size morphing is not significantly slower than the (idealistic but impractical) gold
standard of absolute spatial stability.
In future work we will conduct a longitudinal study of user performance with a
complete tool implementing the principles derived and tested in this paper.
References
1. Hutchings, D., Smith, G., Meyers, B., Czerwinski, M., Robertson, G.: Display Space Usage
and Window Management Operation Comparisons between Single Monitor and Multiple
Monitor Users. In: Proc. AVI 2004, pp. 32–39. ACM Press, New York (2004)
2. Smith, G., Baudisch, P., Robertson, G., Czerwinski, M., Meyers, B., Robbins, D., Horvitz,
E., Andrews, D. GroupBar: The TaskBar Evolved. In: Proc. OzCHI 2003, pp. 34–43 (2003)
3. Henderson, A., Card, S.: Rooms: The use of multiple virtual workspaces to reduce space
contention in a window-based graphical user interface. ACM Transactions on Graphics 5,
3, 211–243 (1986)
4. Robertson, G., Horvitz, E., Czerwinski, M., Baudisch, P., Hutchings, D., Meyers, B.,
Robbins, D., Smith, G.: Scalable Fabric: Flexible Task Management. In: Proc. AVI 2004,
pp. 85–89. ACM Press, New York (2004)
5. Kumar, M., Paepcke, A., Winograd, T.: EyeExposé: Switching Applications with Your
Eyes, Stanford University, 2007. hci.stanford.edu/cstr/reports/2007-02.pdf
6. Grudin, J.: Partitioning Digital Worlds: Focal and Peripheral Awareness in Multiple
Monitor Use. In: Proc. CHI 2001, pp. 458–465. ACM Press, New York (2001)
7. Oliver, N., Smith, G., Thakkar, C., Surendran, A.: SWISH: Semantic Analysis of Window
Titles and Switching History. In: Proc. IUI 2006, pp. 194–201. ACM Press, New York
(2006)
8. Kaptelinin, V.: UMEA: Translating Interaction Histories into Project Contexts. In: Proc.
CHI 2003, pp. 353–360. ACM Press, New York (2003)
9. Dragunov, A., Dietterich, T., Johnsrude, K., McLaughlin, M., Li, L., Herlocker, J.:
TaskTracer: A Desktop Environment to Support Multi-tasking Knowledge Workers. In:
Proc. IUI 2005, pp. 75–82. ACM Press, New York (2005)
10. Oliver, N., Czerwinski, M., Smith, G., Roomp, K.: RelAltTab: Assisting users in switching
windows. In: Proc. IUI 2006, pp. 385–388. ACM Press, New York (2006)
11. Bernstein, M., Shrager, J., Winograd, T.: Taskposé: Exploring Fluid Boundaries in an
Associative Window Visualization. In: Proc. UIST 2008, pp. 231–234. ACM, New York
(2008)
12. Shneiderman, B.: Direct Manipulation for Comprehensible, Predictable, and Controllable
User Interfaces. In: IUI 1997, pp. 33–39. ACM Press, New York (1997)
13. Tashman, C. WindowScape: A Task Oriented Window Manager. In: Proc. UIST 2006, pp.
77–80. ACM Press, New York (2006)
14. Kandogan, E., Shneiderman, B.: Elastic Windows: Improved Spatial Layout and Rapid
Multiple Window Operations. In: Proc. AVI 1996, pp. 29–38. ACM Press, New York
(1996)
15. Gaylin, K.: How are Windows Used? Some Notes on Creating an Empirically-Based
Windowing Benchmark Task. In: Proc. CHI 1986, pp. 96–100. ACM Press, New York
(1986)
16. Mackinlay, J., Royer, C.: Log-based Longitudinal Study Finds Window Thrashing, Palo
Alto Research Center, UIS-2004-06 (2004)
17. Zipf, G.: Human Behavior and the Principle of Least Effort: An Introduction to Human
Ecology. Addison-Wesley, Reading, Mass. (1949)
18. Juran, J.: Quality Control Handbook. McGraw-Hill, New York (1951)
19. Cockburn, A., Gutwin, C., Greenberg, S.: A Predictive Model of Menu Performance. In:
Proc. CHI 2007, pp. 627–636. ACM Press, New York (2007)
20. Tauscher, L., Greenberg, S: How People Revisit Web Pages: Empirical Findings and
Implications for the Design of History Systems. IJ Hum. Comp. Stud, 47, 1 (1997), 97-138.
21. Hansen, W.: User Engineering Principles for Interactive Systems. In: Barstow, D., Shrobe,
H., Sandewall, E. (eds.) Interactive Programming Environments, pp. 288–299. McGraw-
Hill, New York (1984)
22. Czerwinski, M., van Dantzich, M., Robertson, G., Hoffman, H.: The Contribution of
Thumbnail Image, Mouse-Over Text and Spatial Location Memory to Web Page Retrieval
in 3D. In: Proc. INTERACT 1999, pp. 163–170. Springer (1999)
23. Bruls, M., Huizing, K., van Wijk, J.: Squarified Treemaps. In: Proc. Eurographics and
IEEE TCVG Symposium on Visualization, pp. 33–42. IEEE Press (2000)
24. Tu, Y., Shen, H.: Visualizing Changes of Hierarchical Data Using Treemaps. IEEE Trans.
on Visualization and Computer Graphics 13, 6, 1286–1293 (2007)
